See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/229047219

Image Watermarking Using the Empirical Mode Decomposition

Article

CITATIONS READS
7 275

4 authors, including:

Sabri Abdelouahed H. Tairi

Sidi Mohamed Ben Abdellah University Sidi Mohamed Ben Abdellah University
89 PUBLICATIONS 498 CITATIONS 213 PUBLICATIONS 1,185 CITATIONS

SEE PROFILE SEE PROFILE

Abdellah Aarab
Sidi Mohamed Ben Abdellah University
82 PUBLICATIONS 422 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Medical Image Processing Project View project

Road Safety in Morocco View project

All content following this page was uploaded by H. Tairi on 26 April 2014.

The user has requested enhancement of the downloaded file.

 SETIT 2009
5th International Conference: Sciences of Electronic,
Technologies of Information and Telecommunications
March 22-26, 2009 – TUNISIA

Image Watermarking Using the Empirical Mode Decomposition

Abdelouahed SABRI*, Mohamed KAROUD*, Hamid TAIRI**
and Aabdelah AARAB*
* LESSI, Department of physics,
Faculty of Science, Dhar El mahraz BP 1796 FES. Morocco
abdelouahed.sabri@gmail.com
aaroud_moh@yahoo.fr
aarab_abdellah@yahoo.fr

** LIIAN Department of maths and informatics

Faculty of Science, Dhar El mahraz BP 1796 FES. Morocco
htairi@yahoo.fr

Abstract: With the increasing amount of digital data sharing tools due to the growth of the internet; thing that increased
the availability of images to the internet users; frauds are multiplied which accentuated the need to find robust marking
techniques for digital data protection. Researches prove that watermarking is a valid solution for such a need. Here, we
propose a robust watermarking technique based on the BEMD (Bidimensional Empirical Mode Decomposition) where
an additive watermarking scheme is applied to each IMF (Intrinsic Mode Function) obtained by the BEMD and the
watermark is a binary matrix differently ponderated for each one of the four first IMFs. The use of the BEMD in
watermarking is motivated by the fact that it has better quality than Fourier, Wavelet and other decomposition
techniques in extracting intrinsic components because of its fully data driven property, and this decomposition is also
proven as a very powerful tool for multi-scale analysis of non stationary and non linear signals and also by the
characteristics of the IMF. All this mean that, our method can be applied to any image. Experimentation has ensured the
mark invisibility, the facility of detection of the mark and the robustness against different kinds of attacks.
Key words: BEMD, JPEG, noise, private watermarking scheme, robustness, Watermarking, Watermark invisibility.

The basis idea of our method is to insert a different

INTRODUCTION
To control unauthorized copying and illegal
distribution of multimedia data, a digital watermarking
method at first consists in embedding an imperceptible
watermark and it should be robust against attacks.
Robustness means that the watermark should be
detectable as long as the quality of the content itself is
not rendered useless. Attacks to watermarked digital
images can be accidental as in image processing like in
denoising, lossy compression … or intentional which
aim to distort the watermark or to find the secret key
used to generate the watermark. A watermarking
scheme can be private (not blind) or public (blind),
depending on the need or not of the original image.
Here we use the private scheme where the original
image is used in the detection process. 
A variety of watermarking techniques has been
proposed recently in the literature [Red 05]-[San 07]-
[Yon 08]-[Che 07], meanwhile each method proves its
efficiency against special attacks and none of them is
totally effective against malicious attacks.
mark into the each intrinsic mode after decomposing
the image by the EMD (Empirical Mode
Decomposition). The concept of the empirical mode
decomposition is to expand any oscillatory signal into a
set of functions defined by the signal it self, Intrinsic
Mode Functions (IMF), the process is called sifting
process. An IMF is characterized by some specific
properties; the first is that the number of zero crossings
and the number of extrema points is equal or differs by
only one, and the second is the envelopes defined by the
local maxima and minima, respectively, are locally
symmetric around the envelope mean. This scheme of
insertion has the characteristic to provide a good
performance of detection of the mark against attacks
because of the redundancy of the watermark in the
watermarked image (the mark is inserted the number of
the IMF concerned time in the image).
1. The Empirical Mode Decomposition
EMD is an adaptive decomposition of data,
introduced by Huang [Hua 98] for one dimensional
signal and extended to image [Nun 03] after that.

-1-
SETIT2009
Researches proves that the EMD is as better quality
than Fourier, wavelet and other decomposition
techniques in extracting intrinsic components of
textures and image compression because of its fully
data driven property. This decomposition is also proven
as a very powerful tool for multi-scale analysis of non
stationary and non linear signals.
The BEMD is a highly adaptive decomposition. It is
based on the characterization of the image through its
decomposition in intrinsic mode function (IMF) where
the image can be decomposed into a redundant set of
signals denoted IMF and a residue, adding all the IMF’s
together with the residue reconstructs the original image
without information loss or distortion. An IMF is
characterized by some specific properties:
• The number of zero crossing and the number of
extrema points is equal or differs only by one,
• It has a zero local mean.
Given a 2D signal we can describe the principal
of the empirical mode decomposition as follows:
1. Initialisation:
number of IMF);
(the residual) and j (index
2. Extract the jth IMF:
a. Initialise h0 = rj −1 , i = 1 .
b. Extract local minima and maxima of h . such a way to increase (or decrease) by a unit the
i −1
c. Compute upper envelope and lower envelope
functions xi −1 and yi −1 by interpolating respectively
local minima and local maxima of h . of the pixels of the type Bi . In such way there will
i −1
d. Compute mi −1 = ( xi −1 + yi −1 ) / 2 , the mean envelope.
e. Update hi = hi −1 − mi −1 and i = i + 1 .
f. Calculate stopping criterion
1 k ⎡ (hi ( j −1) ( k ) − hij ( k )) ⎤
2
SD = ∑⎢ ⎥ ⎪ 0
N k =0 ⎢⎣ hi2( j −1) ( k ) ⎦⎥
g. Decision: Repeat steps (b) to (f) until SDi =< ξ ,
and then put d i = hi ( jht IMF ) .
3. Update residual rij = rj −1 − m j .
4. Repeat steps 1-3 with j := j + 1 until the number of
extrema in rj is less than 2.
When the decomposition is complete, we can write
the signal in the following form:
K
m( n) = ∑ d k ( n) + r ( n ) , K ∈ N *
k =1
-2-
The decomposition of images of great size using the
BEMD can necessitates an important calculation time,
so to overcome this problem we can accelerate the
process and minimize the computational time by the use
of the LBBEMD (Lapped Block Bidimensional
Empirical Mode Decomposition) [Kar 06].
2. Watermark embedding using the BEMD
As a first application of image watermarking using
the IMF after decomposition by BEMD, we will use the
“Patchwork” algorithm; proposed by Bender and Al.
[Ben 95]. We will apply the additive watermarking
scheme to each IMF obtained by the BEMD where the
watermark is a binary matrix in {- 1, 1} differently
ponderated for each one of the four first IMFi modes,
i=1… 4, and at the end watermarked image is
reconstructed by adding all the watermarked IMF and
the residue (Figure 1.). The watermark insertion scheme
is as follow:
Step.1 Decompose the original image to the fourth first
modes and residue:
n
I= ∑ IMF + residue
i =1
i with n = 4
Step.2 for each IMF, using a secret private key to select
randomly a sequence of n couples of
pixels ( Ai , Bi ) in the IMF to be modified lightly in
intensity value of the pixels of the type Ai and to
decrease (or increase) by a unit the intensity value
be a unique private key for each IMF.
Step.3 Embedding in each IMF the equivalent mark:
IMFwi = IMFi + α i × Ri with i = i...4
⎧ 1 If pixels are in Ai
⎪
and R i = ⎨− 1 If pixels are in Bi
⎩ Others
IMF is the original intrinsic mode function,
IMFw is the watermarked intrinsic mode
function, α is a ponderation and R is the binary
watermark matrix
Step.4 Building the watermarked image by adding all
the watermarked IMF and the residue of the
decomposition of the original image by the BEMD:
n
IW = ∑ IMFw
i =1
i + residue with n = 4
SETIT2009

Figure 1. Diagram of the image watermarking method using BEMD

Step.3 Extraction of the watermark from each IMF:

3. Watermark detection using the BEMD
The watermark detection is done by operation dual of IMFi* = IMFwi + α i × Ri with i = i...4
the one used in image watermarking process (Figure.2.). If pixels are in Ai
The watermark detection scheme is as follow: ⎧− 1
⎪
and Ri = ⎨ 1 If pixels are in Bi
Step.1 Decompose the watermarked image to the fourth
⎪ 0
first modes and residue: ⎩ Others
n
IW = ∑ IMFw + residue
i with n = 4 IMFw is the watermarked intrinsic mode function,
1 IMF * is the intrinsic mode function after extracting
Step.2 For each MFw , and using the secret private key;
the watermark, α is a ponderation and R is the
used in the referencing stage in the watermarking
binary watermark matrix
process; select the sequence of n couples pixels (Ai,
Step.4 Verifying the existence of the watermark by a
Bi) in the MFw to be modified lightly in such a way threshold on the result of correlation between each
to increase (or decrease) by a unit the intensity value
of the pixels of the type Ai and to decrease (or IMFi* and IMFi and between the original image
increase) by a unit the intensity value of the pixels of and the image after extraction of the watermark
the type Bi.

-3-
SETIT2009
Figure 2. Diagram of watermark detection method using BEMD
4. Experimental results
4.1 Watermark embedding and detection without
attacks
A first application is to test our method for a
grayscale image (Lena) of 512 x 512 and try to extract
the watermark without any attack on the watermarked
image.
4.1.1 Embedding
Figure 3. Watermark embedding: a) Original image, b)
Watermarked image, c) difference between original
image and watermarked image.
In Figure.3. we presented in a) the image before
insertion of the watermark (original image), in b) the
watermarked image is presented to show the invisibility
of the watermark in the watermarked image and at the
end in c) the difference (the watermark) between the
original image and the watermarked image is presented.
It is clear that the first characteristic of a
watermarking method which is the invisibility of the
watermark is respected in our method.
-4-
4.1.2 Detection
We presented in Figure.4. the results of the detection
of the watermark in the watermarked image. So, to
verify the existence of the mark in the watermarked
image, a measure of correlation between the
watermarked image after extraction of the watermark
and the original image is done.
Figure 4. Detection of the existence of the watermark
without any attack: a) Watermarked image, b) image
after extraction of the watermark, c) correlation
between the watermarked image after extraction of the
watermark and the original image as well as the
correlation between the modes (IMF).
If there are no attacks the detection of the watermark
is done without any problem.
4.2 Robustness of the method against attacks
Certainly once watermarked images are diffused in
the internet can beings victim of attacks and arbitrary
transformed. If these attacks do not much degrade the
watermarked image, a robust watermarking method
should detect the watermark as long as the watermarked
image quality is not rendered useless. So, to test the
robustness of our method against attacks, we will
present some examples of attacks/detection.
SETIT2009
4.2.1 Robustness against white noise attack
We can suppose that for image with moderate
signal-to-noise ratio, the noise is of lower amplitude
than the extrema points. We will apply to the
watermarked image a Gaussian white additive noise
with zero mean. We can suppose that for image with
moderate signal-to-noise ratio, the noise is of lower
amplitude than the extrema points. So, after
decomposing the watermarked image, we can reduce the
noise by setting the extrema points with low amplitude
to zero in the first IMF; the remaining values represent
only significant extrema points. That means that the
problem of noise attacks on the watermarked images
could be treated at the level of the first mode without
influencing a lot the other modes.
Figure 5. a) Watermarked Image with white noise, b)
Watermarked image after reducing the noise, c)
correlation between the watermarked image after
extraction of the watermark and the original image as
well as the correlation between the modes (IMF).
The correlation results show that the attack by the
white noise has been felt on the first IMF and did not
have big impacts on the other modes. Thus, we can say
that our method is robust against an attack of against
white noise.
4.2.2 Robustness against JPEG compression attack
Image compression consists in transmitting only the
significant components, and since the watermark is
invisible in the watermarked image it will be seen in the
compression process as non useful information. To
remedy this problem Cox and al [Cox 97] proposed to
insert the watermark in significant places in the image.
The idea to make our method robust to the JPEG
compression is to modify the DCT coefficients [Koc 95]
of each IMF en the JPEG process as follow:
• Split Each IMF into non-overlapping 8x8 blocks.
• Take the two-dimensional DCT of each block.
• Scanning the DCT component matrix in zigzag pattern
to keep only significant components located in the
upper left corner.
• Finally, the modifications are done on the low
frequency coefficient quantified triplet. In this way we
are sure that the watermark will be not removed form
the image and also it will be invisible.
The existence of the watermark is checked by
comparing the original image and the decompressed
image watermarked image after extracting the
watermark.
-5-
Figure 6. a)watermarked image JPEG compressed
using quality 50%, b) Image after watermark extraction,
c) Correlation between the two images
In Figure.6. we presented the watermarked image
with quality factor equal to 50%, the image after
extracting the watermark and the correlation between
the two images and the IMF. The correlation results
show that the watermark exists in the watermarked
image after Jpeg compression and decompression.
To evaluate our method for different JPEG quality
factor, we tested to verify the existence of the
watermark in the JPEG compressed watermarked image
with different quality factor. Table 1. presents the
results of correlation for different JPEG quality factor.
JPEG quality 20 40 50 70 80 90 100
Correlation 0.5460 0.7890 0.9968 0.9988 0.9992 0.9995 1
Table 1. Correlation results for different JPEG quality
factor (%)
The results obtained are very encouraging in terms
of visual; the PSNR equal to 39.50dB; and also the
correlation result between the compressed image after
watermark extraction and the original image is very
encouraging.
5. Conclusion
In this paper we presented a new image
watermarking based on the BEMD decomposition. This
method presents an important asset because of the
redundancy of the watermark in the watermarked image
(four different watermark are present in the
watermarked image), thing which gives us four times
more robustness with respect to the methods aiming to
extract the watermark. Although, we did not present the
robustness of the method with respect to all the attacks
but seen the results presented here our method presents
an excellent behavior against noise and JPEG
compressions attacks which lets us say that it is very
promising images watermarking method.
REFERENCES
[Ben 95] W. Bender, D. Gruhl & N. Morimoto. Techniques for
data hiding. Proceedings of the SPIE, 1995.
[Che 07] C. Chemak, M.S. Bouhlel, and J.-C. Lapayre, “A
new scheme of image Watermarking based on 5/3 Wavelet
decomposition and Turbo-Code,” WSEAS Transaction on
Biology and Biomedicine, April 2007, pp. 45-52.
 SETIT2009

[Cox 97] I.J. Cox, J. Kilian, F.T. Leighton, and T. Shamoon.
Secure spread spectrum watermarking for multimedia.
IEEE Transactions on Image Processing, 6(12)
:16731687, 1997.
[Hua 98] NE. Huang, Z. Shen, SR. Long, MC. Wu, HH. Shih,
Q. Zheng, N-C. Yen, CC. Tung, HH. Liu(1998), The
empirical mode decomposition and the Hilbert spectrum
For non-linear and non-stationary time series analysis,
Proc R Soc Land Ser A 454:903-995.
[Kar 06] M. Karoud, M. A. Sabri, S. J.Andaloussi, H. Tairi and
A. Aarab, Block Image Analysis using Empirical Mode
Decomposition, journal WSEAS Transactions on
Computers. Issue 12, Volume 5, December 2006, P : 2903-
2911.
[Koc 95] E. Koch and J. Zhao. Towards robust and hidden
image copyright labeling. In Proc. of 1995 IEEE Workshop
on Nonlinear Signal and Image Processing, pages 452455,
Halkidiki, Greece, June 1995.
[Nun 03] J. C. Nunes, Y. Bouaoune, E. Delechelle, O. Niang,
Ph. Bunel, Image analysis by bidimensional empirical
mode decomposition, Image and Vision Computing
(2003).
[Red 05] A. Reddy Adhipathi, B.N. Chatterji, A new wavelet
based logo-watermarking scheme, Pattern Recognition
Letters 26 (2005) 1019–1027.
[San 07] Santa Agreste, Guido Andaloro, A new approach to
pre-processing digital image for wavelet-based watermark,
Journal of Computational and Applied Mathematics 210
(2007) 13 – 21
[Yon 08] Yong-Gang Fu, Rui-Min Shen, Color image
watermarking scheme based on linear discriminant
analysis, Computer Standards & Interfaces 30 (2008) 115–
120

-6-
View publication stats